Dynamic Fusion of YOLOv8 and Multimodal Features for Small Object Detection in Inland Waters

Inland water environments pose unique challenges for small object detection due to weak textures, low contrast, occlusions, and background interference. Traditional visual detection methods such as YOLO and Faster R-CNN struggle to maintain sufficient precision under these conditions. This paper presents a dynamic fusion detection system that integrates YOLOv8 with multimodal features including RGB images, millimeter-wave radar (RDM), and sparse point clouds. The system adaptively adjusts fusion weights based on object size and confidence, enhancing robustness for small object detection. A weak-label-based perception validity prediction task is introduced to estimate detection validity and provide spatial localization proposals, addressing the all-positive sample constraint of the FloW-RI dataset. Experiments show that the proposed system outperforms conventional methods, achieving an F1 score of 0.95 and Recall up to 0.99. Ablation studies and visualizations confirm the contributions of the YOLO-assisted module, cross-modal attention mechanism, and dynamic fusion strategy. Future work will explore adaptability to occlusions, improved perception validity estimation, and real-time deployment.